• Home
• /
• Opinion

ChatGPT's lessons for economic development

To learn, AI algorithms require training, which can be achieved through two main approaches: supervised and unsupervised learning

Spoiler alert: I am not going to talk about how ChatGPT responds when prompted about economic-development strategies. It basically regurgitates reasonable, but mediocre ideas that it has seen in its training set. But ChatGPT’s design, which has given it far greater capabilities than its creators anticipated, offers a valuable lesson for tackling the complexities of economic development.

For more than a decade, deep neural networks (DNNs) have outperformed all other artificial-intelligence technologies, driving significant advances in computer vision, speech recognition, and translation. The emergence of generative AI chatbots like ChatGPT continue this trend.

To learn, AI algorithms require training, which can be achieved through two main approaches: supervised and unsupervised learning. In supervised learning, humans provide the computer with a set of labelled pictures such as “dog,” “cat”, “hamburger”, “car” and so on. The algorithm is then tested to see how well it predicts the labels associated with images it has not yet seen.

The problem with the supervised approach is that it requires humans to go through the tedious process of manually labelling every picture. By contrast, unsupervised learning does not rely on labelled data. But the absence of labels raises the question of what the algorithm is supposed to learn. To address this, ChatGPT trains the algorithm simply to predict the next word of the text that is used to train it.

Predicting the next word may seem like a trivial task, akin to the auto-complete function in Google Search. But ChatGPT’s model allows it to perform highly complicated tasks, such as passing the bar exam with a better score than most high-performing law students.

The key to such feats lies in the impressive power of this simple learning process. In order to predict the next word, the algorithm is forced to develop a nuanced understanding of context, grammar, syntax, style, and much more. The level of sophistication it achieved surprised everybody, including its designers. DNNs proved capable of functioning much better without trying to incorporate into learning language models the theories that linguists had developed for decades.

The lesson for economic development is that policymakers should focus on a task that may seem mundane, provided that to excel at it, they will indirectly be forced to learn much more intricate development challenges.

By contrast, the prevailing approach in the field of development economics has been to distinguish between proximate causes and deeper determinants of growth and to focus on the latter. This approach is analogous to saying, “Instead of trying to predict the next word, understand the context and meaning of the entire book”.

In their 2012 book Why Nations Fail, for example, Daron Acemoglu and James A. Robinson argue that institutions, by affecting the structure of incentives in society, are the ultimate determinant of economic outcomes. Brown University economist Oded Galor has taken a different approach, emphasising the complex demographic and technological transformations that brought humanity out of the Malthusian equilibrium and led to longer life expectancy, lower fertility rates, and higher investment in education. Together, these trends boosted women’s participation in the labor force and increased the availability of skills needed to sustain technology adoption and economic growth.

But do these theories match the facts? Over the past four decades, the developing world has indeed undergone many of the radical transformations that Galor described. As the late physician Hans Rosling observed, the gaps between developing and developed countries in life expectancy, infant mortality, fertility, education, university enrollment, female labour-force participation, and urbanisation have all narrowed sharply. Reasoning à la Acemoglu and Robinson, developing countries’ institutions could not be all that bad if they were able to deliver progress on so many fronts. In Galor’s framework, progress on all these fronts should explain why developing countries caught up so much with the developed world in terms of income.

Except that they did not: the median country is no closer to US income levels than it was four decades ago. How is it possible that the narrowing gaps in education, health, urbanisation, and female empowerment failed to narrow the income gap as well? Why hasn’t progress in the supposed deeper determinants delivered the goods?

To make sense of this puzzling outcome, economists invoke a widening technological gap. More than an explanation, this is a mathematical necessity: if more inputs do not generate more output, something must be making inputs less effective.

To explain this unexpected outcome, it is useful to note that the few countries that did manage to catch up share two distinctive features: their exports grew much faster than their GDP, and they diversified their exports by shifting toward more complex goods.

To achieve this feat, these successful countries must have adopted and adapted better technologies, adjusted the provision of public goods and their institutions to support emerging industries, and reduced inefficiencies and costs by increasing productivity and training workers. In that process, they may have fixed a bunch of other problems.

A ChatGPT-inspired development strategy would focus on a simple goal: to improve the competitiveness, diversity, and complexity of exports. Figuring out how to do this would force policymakers to learn how to do important things, just like predicting the next word enabled ChatGPT to learn context, grammar, syntax, and style.

Like early AI programmers who were sidetracked by linguists and their convoluted theories, policymakers have been distracted by too many objectives, such as the 17 United Nations Sustainable Development Goals. But applying the ChatGPT approach to economic development could simplify things: just as the language model tries to predict only the next word, policymakers could try to focus on facilitating the next export, as successful countries seem to have done. While this may seem like a small step, it could lead to surprisingly significant results.— Project Syndicate

Ricardo Hausmann, a former minister of planning of Venezuela and former chief economist at the Inter-American Development Bank, is a professor at Harvard University’s John F. Kennedy School of Government and Director of the Harvard Growth Lab.